fix(core): preserve the first terminal failure

[kitlangton]

What

Preserve the first causal assistant failure while concurrent tool fibers settle, and remove malformed-input event fields that no production consumer reads.

Before / After

Before: A provider failure could be recorded first, then overwritten by a later tool-output persistence failure because every terminal cleanup path forced failAssistant(..., true). Durable history reported the cleanup error instead of the provider error that ended the step.

After: failAssistant is write-once. Later interruption, persistence, or missing-hosted-result cleanup still settles tools, but cannot replace the first terminal assistant failure.

How

• packages/core/src/session/runner/llm.ts stops forcing replacement from terminal cleanup paths.
• packages/core/src/session/runner/publish-llm-event.ts removes the replacement parameter and keeps the first stepFailure.
• packages/core/test/session-runner.test.ts covers a provider failure racing with local tool-output persistence failure.
• packages/ai and the AI SDK adapter remove unused ToolInputError.message and providerMetadata fields while retaining the authoritative raw malformed input.
• to-llm-message.ts reuses the existing provider-metadata predicate instead of spelling the same condition twice.

Scope

• Provider-executed malformed calls remain terminal.
• Local malformed calls remain ordinary failed, unexecuted tool results and continue through the normal agent Step policy.
• Failed-Step snapshot and file authority remains intact for revert planning.
• This does not add a separate malformed-call retry budget.

Testing

• bun run test in packages/ai: 343 passed, 29 skipped.
• bun run test -- test/session-runner.test.ts test/session-runner-message.test.ts test/session-runner-tool-events.test.ts test/aisdk.test.ts in packages/core: 169 passed.
• bun turbo typecheck --concurrency=3: 32 tasks passed; the push hook repeated this successfully.
• bun run lint: completed with existing repository warnings and no new errors.
• Full Core run: 1,347 passed and 6 skipped. Three unrelated environment/timing cases failed: configured execute, user-level cmux plugins, and the existing location flush timing test.
• git diff --check: passed.

Flow

flowchart TD
  P[Provider failure] --> F[failAssistant]
  F --> S{stepFailure set?}
  S -- no --> R[Record provider failure]
  T[Concurrent tool settlement failure] --> F
  S -- yes --> K[Keep first failure]
  R --> E[Publish failed Step]
  K --> E

Loading